DeepSeek Harness plugin

dsh-computer-use-windows

Windows computer-use for DeepSeek Harness: window-bound screenshots, OCR, and verified clicks.

Jump to install

Source facts

Repository
Altairpaca/dsh-computer-use-windows
Latest update
Aug 14, 2026
Category
Tools & Capabilities
GitHub stars
1
Format
plugin
Catalog evidence
Upstream dsh.bundle evidence
Evidence path
package.json#dsh.bundle
Checked against
0.1.0-rc.8
Upstream check date
2026-08-20

This evidence comes from the upstream catalog. This site has not installed, run, or security-reviewed the plugin.

Install

Start with a prompt that asks an agent to review the GitHub repository and source. Switch to the command if you want to install it yourself.

Copy this prompt into DSH, Codex, or another agent and ask it to review the GitHub repository and source first.

Do not install or run any commands yet. Read this plugin's GitHub repository, README, and relevant source code. Then answer the questions below clearly and directly so I can decide whether it fits my needs:

1. What is this plugin, and what problem does it solve?
2. Who is it for, and what are its typical use cases?
3. How is it used after installation? Include one minimal example.
4. What known limitations or privacy, security, compatibility, or maintenance risks does it have?
5. Give a clear recommendation: recommend, conditionally recommend, or do not recommend, with reasons.

Distinguish statements documented by the repository, inferences from source code, and unknowns. If evidence is insufficient, say so explicitly. Do not guess or simply repeat the README.

GitHub: https://github.com/Altairpaca/dsh-computer-use-windows
Plugin: dsh-computer-use-windows
Author: Altairpaca

Check the source files

Read the README and other files from this plugin directory before installing.

File explorer3 files
README.mdSource · read only

DSH Computer Use (Windows)

> Windows Computer Use for DeepSeek Harness: > window-bound screenshots, robust OCR, verified clicks, pure-OCR mode, and pluggable vision models. > > 为 DeepSeek Harness 提供的 Windows 电脑控制插件:窗口绑定截图、健壮 OCR、带验证闭环的点击、 > 纯 OCR 模式、可自由配置的视觉模型。填补 DSH 生态 Windows computer use 的空白。

![License: MIT](LICENSE) !Platform !PowerShell !DeepSeek Harness

为什么存在

DSH 生态里已有 Anionex/dsh-computer-use(macOS, Accessibility-first)与一批 vision 插件,但 Windows 上没有语义优先的 computer use 插件。 本项目把 2026-08-15 真实实验(操作保险"云助理"客户端提取客户数据)的教训固化为插件:

实验中最致命的三个问题,全部来自"盲坐标"工作流:

1. VLM/OCR 估坐标 → 点击系统性偏移("识别偏上、行为偏下"),点错入口; 2. 全屏截图混入其他窗口 → OCR 结果被无关文字污染,坐标张冠李戴; 3. 无验证闭环 → 每次点击都是盲试,靠人工盯着纠正。

本项目用四条原则解决它们:

  • 窗口绑定:所有截图/OCR/点击都在一个目标窗口的坐标系内,点击时自动回加窗口偏移;
  • 文字即坐标:提供 click_text,用 OCR 定位文字再点击——模型不再需要猜坐标;
  • 验证闭环:每次点击后自动截图 + OCR 校验预期状态,失败按偏移网格重试,返回证据;
  • 可配置感知:视觉模型可插拔(任意 OpenAI 兼容端点),也可以完全关闭——纯 OCR 模式

下全部操作不依赖任何视觉模型,零外部 API。

功能特性

能力说明
computer_screenshot全屏或目标窗口截图,返回窗口 rect 与坐标空间信息
computer_ocrWinRT OCR(中文/英文),逐词坐标,窗口外词自动过滤,模糊匹配查询
computer_click_text按文字点击:OCR 定位 → 点击 → 验证 → 失败重试(偏移网格)
computer_mouse / computer_keyboard绝对坐标注入、拖拽、滚轮、剪贴板输入
computer_window窗口枚举/聚焦/定位,句柄动态解析,前台校验
computer_use_run批量动作一次调用(点击+输入+验证)
computer_vision(可选)可插拔视觉:opencode-go / 任意 OpenAI 兼容端点 / 关闭
computer_calibrateDPI/残差校准,校准结果持久化,后续点击自动修正

两种工作模式(Config 自由)

{
  "vision": {
    "enabled": false            // ← 纯 OCR 模式:不调用任何视觉模型
  }
}
{
  "vision": {
    "enabled": true,
    "provider": "openai-compatible",   // opencode-go | openai-compatible | none
    "base_url": "https://your-vlm.example.com/v1",
    "api_key_env": "MY_VLM_KEY",       // 从环境变量/DSH 凭据库取 key,不进配置
    "model": "your-vlm-model"
  }
}

完整配置见 [docs/design.zh.md](docs/design.zh.md)(config schema、每个字段的语义与默认值)。

快速开始

> 当前处于 v0 参考实现阶段(调研 + 设计 + helper 实现已完成,bundle 集成验证中)。 > 已安装 DSH(Web profile)与 PowerShell 7.4+。

# 1. 克隆
git clone https://github.com/Altairpaca/dsh-computer-use-windows
# 2. 体检(可选)
./scripts/check-health.ps1
# 3. 直接使用 helper(无需安装插件即可体验)
$env:CU_ARGS = '{"cmd":"screen"}'
& ./helper/cu.ps1
# 4. 把 helper 安装为 DSH bundle(待 v1 完成)
dsh plugin add dsh-computer-use-windows   # 规划中

项目状态与路线图

阶段内容状态
0. 复盘实验问题根因分析(坐标/验证/窗口)✅ 完成([docs/experiment-findings.zh.md](docs/experiment-findings.zh.md))
1. 调研跨平台 computer use 方案调研(Claude/OpenAI/Gemini/UFO/OmniParser/UI-TARS…)📋 计划见 [docs/research-plan.zh.md](docs/research-plan.zh.md)
2. 设计架构、健壮性、config 自由度、OCR 健壮性✅ 完成([docs/design.zh.md](docs/design.zh.md))
3. 参考实现cu.ps1 v2(窗口绑定/click_text/验证闭环/配置化)helper/cu.ps1
4. Bundle 化DSH 插件封装(tools 注册 + SKILL + 一键安装)🚧 plugins/ 骨架,待调研结论固化
5. 发布GitHub Discussion、awesome-dsh-plugin 收录📋 草稿见 [docs/discussion-post.md](docs/discussion-post.md)

文档

  • [docs/research-plan.zh.md](docs/research-plan.zh.md) — 跨平台调研计划(问题清单/平台矩阵/验证方法)
  • [docs/design.zh.md](docs/design.zh.md) — 架构设计、健壮性、config schema、OCR 健壮性
  • [docs/experiment-findings.zh.md](docs/experiment-findings.zh.md) — 实验复盘与根因
  • [skills/computer-use-windows/SKILL.md](skills/computer-use-windows/SKILL.md) — 模型侧使用技能

安全声明

  • 本插件通过 user32 注入鼠标键盘,可操作系统任意窗口;默认只在目标窗口内点击,

但仍建议在隔离环境/授权范围内使用;

  • vision 调用会把截图发送到你配置的视觉端点(自备端点或纯 OCR 模式可避免外发);
  • API key 只从环境变量或 DSH 凭据库读取,不进入配置文件与代码。

许可证

[MIT](LICENSE) © 2026 Altair Li

---

灵感与致谢:Anionex/dsh-computer-use(macOS 语义优先设计)、 ezpzai/codex-computer-use-windows(Windows 配方)、 ysr666/dsh-vision-router(grounding 工具族)。